1 //===- BreakCriticalEdges.cpp - Critical Edge Elimination Pass ------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // BreakCriticalEdges pass - Break all of the critical edges in the CFG by 10 // inserting a dummy basic block. This pass may be "required" by passes that 11 // cannot deal with critical edges. For this usage, the structure type is 12 // forward declared. This pass obviously invalidates the CFG, but can update 13 // dominator trees. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "llvm/Transforms/Utils/BreakCriticalEdges.h" 18 #include "llvm/ADT/SetVector.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/Analysis/BlockFrequencyInfo.h" 22 #include "llvm/Analysis/BranchProbabilityInfo.h" 23 #include "llvm/Analysis/CFG.h" 24 #include "llvm/Analysis/LoopInfo.h" 25 #include "llvm/Analysis/MemorySSAUpdater.h" 26 #include "llvm/Analysis/PostDominators.h" 27 #include "llvm/IR/CFG.h" 28 #include "llvm/IR/Dominators.h" 29 #include "llvm/IR/Instructions.h" 30 #include "llvm/InitializePasses.h" 31 #include "llvm/Transforms/Utils.h" 32 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 33 #include "llvm/Transforms/Utils/Cloning.h" 34 #include "llvm/Transforms/Utils/ValueMapper.h" 35 using namespace llvm; 36 37 #define DEBUG_TYPE "break-crit-edges" 38 39 STATISTIC(NumBroken, "Number of blocks inserted"); 40 41 namespace { 42 struct BreakCriticalEdges : public FunctionPass { 43 static char ID; // Pass identification, replacement for typeid 44 BreakCriticalEdges() : FunctionPass(ID) { 45 initializeBreakCriticalEdgesPass(*PassRegistry::getPassRegistry()); 46 } 47 48 bool runOnFunction(Function &F) override { 49 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 50 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr; 51 52 auto *PDTWP = getAnalysisIfAvailable<PostDominatorTreeWrapperPass>(); 53 auto *PDT = PDTWP ? &PDTWP->getPostDomTree() : nullptr; 54 55 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); 56 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr; 57 unsigned N = 58 SplitAllCriticalEdges(F, CriticalEdgeSplittingOptions(DT, LI, nullptr, PDT)); 59 NumBroken += N; 60 return N > 0; 61 } 62 63 void getAnalysisUsage(AnalysisUsage &AU) const override { 64 AU.addPreserved<DominatorTreeWrapperPass>(); 65 AU.addPreserved<LoopInfoWrapperPass>(); 66 67 // No loop canonicalization guarantees are broken by this pass. 68 AU.addPreservedID(LoopSimplifyID); 69 } 70 }; 71 } 72 73 char BreakCriticalEdges::ID = 0; 74 INITIALIZE_PASS(BreakCriticalEdges, "break-crit-edges", 75 "Break critical edges in CFG", false, false) 76 77 // Publicly exposed interface to pass... 78 char &llvm::BreakCriticalEdgesID = BreakCriticalEdges::ID; 79 FunctionPass *llvm::createBreakCriticalEdgesPass() { 80 return new BreakCriticalEdges(); 81 } 82 83 PreservedAnalyses BreakCriticalEdgesPass::run(Function &F, 84 FunctionAnalysisManager &AM) { 85 auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(F); 86 auto *LI = AM.getCachedResult<LoopAnalysis>(F); 87 unsigned N = SplitAllCriticalEdges(F, CriticalEdgeSplittingOptions(DT, LI)); 88 NumBroken += N; 89 if (N == 0) 90 return PreservedAnalyses::all(); 91 PreservedAnalyses PA; 92 PA.preserve<DominatorTreeAnalysis>(); 93 PA.preserve<LoopAnalysis>(); 94 return PA; 95 } 96 97 //===----------------------------------------------------------------------===// 98 // Implementation of the external critical edge manipulation functions 99 //===----------------------------------------------------------------------===// 100 101 BasicBlock *llvm::SplitCriticalEdge(Instruction *TI, unsigned SuccNum, 102 const CriticalEdgeSplittingOptions &Options, 103 const Twine &BBName) { 104 if (!isCriticalEdge(TI, SuccNum, Options.MergeIdenticalEdges)) 105 return nullptr; 106 107 return SplitKnownCriticalEdge(TI, SuccNum, Options, BBName); 108 } 109 110 BasicBlock * 111 llvm::SplitKnownCriticalEdge(Instruction *TI, unsigned SuccNum, 112 const CriticalEdgeSplittingOptions &Options, 113 const Twine &BBName) { 114 assert(!isa<IndirectBrInst>(TI) && 115 "Cannot split critical edge from IndirectBrInst"); 116 117 BasicBlock *TIBB = TI->getParent(); 118 BasicBlock *DestBB = TI->getSuccessor(SuccNum); 119 120 // Splitting the critical edge to a pad block is non-trivial. Don't do 121 // it in this generic function. 122 if (DestBB->isEHPad()) return nullptr; 123 124 if (Options.IgnoreUnreachableDests && 125 isa<UnreachableInst>(DestBB->getFirstNonPHIOrDbgOrLifetime())) 126 return nullptr; 127 128 auto *LI = Options.LI; 129 SmallVector<BasicBlock *, 4> LoopPreds; 130 // Check if extra modifications will be required to preserve loop-simplify 131 // form after splitting. If it would require splitting blocks with IndirectBr 132 // terminators, bail out if preserving loop-simplify form is requested. 133 if (LI) { 134 if (Loop *TIL = LI->getLoopFor(TIBB)) { 135 136 // The only way that we can break LoopSimplify form by splitting a 137 // critical edge is if after the split there exists some edge from TIL to 138 // DestBB *and* the only edge into DestBB from outside of TIL is that of 139 // NewBB. If the first isn't true, then LoopSimplify still holds, NewBB 140 // is the new exit block and it has no non-loop predecessors. If the 141 // second isn't true, then DestBB was not in LoopSimplify form prior to 142 // the split as it had a non-loop predecessor. In both of these cases, 143 // the predecessor must be directly in TIL, not in a subloop, or again 144 // LoopSimplify doesn't hold. 145 for (BasicBlock *P : predecessors(DestBB)) { 146 if (P == TIBB) 147 continue; // The new block is known. 148 if (LI->getLoopFor(P) != TIL) { 149 // No need to re-simplify, it wasn't to start with. 150 LoopPreds.clear(); 151 break; 152 } 153 LoopPreds.push_back(P); 154 } 155 // Loop-simplify form can be preserved, if we can split all in-loop 156 // predecessors. 157 if (any_of(LoopPreds, [](BasicBlock *Pred) { 158 return isa<IndirectBrInst>(Pred->getTerminator()); 159 })) { 160 if (Options.PreserveLoopSimplify) 161 return nullptr; 162 LoopPreds.clear(); 163 } 164 } 165 } 166 167 // Create a new basic block, linking it into the CFG. 168 BasicBlock *NewBB = nullptr; 169 if (BBName.str() != "") 170 NewBB = BasicBlock::Create(TI->getContext(), BBName); 171 else 172 NewBB = BasicBlock::Create(TI->getContext(), TIBB->getName() + "." + 173 DestBB->getName() + 174 "_crit_edge"); 175 // Create our unconditional branch. 176 BranchInst *NewBI = BranchInst::Create(DestBB, NewBB); 177 NewBI->setDebugLoc(TI->getDebugLoc()); 178 if (auto *LoopMD = TI->getMetadata(LLVMContext::MD_loop)) 179 NewBI->setMetadata(LLVMContext::MD_loop, LoopMD); 180 181 // Insert the block into the function... right after the block TI lives in. 182 Function &F = *TIBB->getParent(); 183 Function::iterator FBBI = TIBB->getIterator(); 184 F.insert(++FBBI, NewBB); 185 186 // Branch to the new block, breaking the edge. 187 TI->setSuccessor(SuccNum, NewBB); 188 189 // If there are any PHI nodes in DestBB, we need to update them so that they 190 // merge incoming values from NewBB instead of from TIBB. 191 { 192 unsigned BBIdx = 0; 193 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) { 194 // We no longer enter through TIBB, now we come in through NewBB. 195 // Revector exactly one entry in the PHI node that used to come from 196 // TIBB to come from NewBB. 197 PHINode *PN = cast<PHINode>(I); 198 199 // Reuse the previous value of BBIdx if it lines up. In cases where we 200 // have multiple phi nodes with *lots* of predecessors, this is a speed 201 // win because we don't have to scan the PHI looking for TIBB. This 202 // happens because the BB list of PHI nodes are usually in the same 203 // order. 204 if (PN->getIncomingBlock(BBIdx) != TIBB) 205 BBIdx = PN->getBasicBlockIndex(TIBB); 206 PN->setIncomingBlock(BBIdx, NewBB); 207 } 208 } 209 210 unsigned NumSplitIdenticalEdges = 1; 211 212 // If there are any other edges from TIBB to DestBB, update those to go 213 // through the split block, making those edges non-critical as well (and 214 // reducing the number of phi entries in the DestBB if relevant). 215 if (Options.MergeIdenticalEdges) { 216 for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) { 217 if (TI->getSuccessor(i) != DestBB) continue; 218 219 // Remove an entry for TIBB from DestBB phi nodes. 220 DestBB->removePredecessor(TIBB, Options.KeepOneInputPHIs); 221 222 // We found another edge to DestBB, go to NewBB instead. 223 TI->setSuccessor(i, NewBB); 224 225 // Record the number of split identical edges to DestBB. 226 NumSplitIdenticalEdges++; 227 } 228 } 229 230 // If we have nothing to update, just return. 231 auto *DT = Options.DT; 232 auto *PDT = Options.PDT; 233 auto *MSSAU = Options.MSSAU; 234 if (MSSAU) 235 MSSAU->wireOldPredecessorsToNewImmediatePredecessor( 236 DestBB, NewBB, {TIBB}, Options.MergeIdenticalEdges); 237 238 if (!DT && !PDT && !LI) 239 return NewBB; 240 241 if (DT || PDT) { 242 // Update the DominatorTree. 243 // ---> NewBB -----\ 244 // / V 245 // TIBB -------\\------> DestBB 246 // 247 // First, inform the DT about the new path from TIBB to DestBB via NewBB, 248 // then delete the old edge from TIBB to DestBB. By doing this in that order 249 // DestBB stays reachable in the DT the whole time and its subtree doesn't 250 // get disconnected. 251 SmallVector<DominatorTree::UpdateType, 3> Updates; 252 Updates.push_back({DominatorTree::Insert, TIBB, NewBB}); 253 Updates.push_back({DominatorTree::Insert, NewBB, DestBB}); 254 if (!llvm::is_contained(successors(TIBB), DestBB)) 255 Updates.push_back({DominatorTree::Delete, TIBB, DestBB}); 256 257 if (DT) 258 DT->applyUpdates(Updates); 259 if (PDT) 260 PDT->applyUpdates(Updates); 261 } 262 263 // Update LoopInfo if it is around. 264 if (LI) { 265 if (Loop *TIL = LI->getLoopFor(TIBB)) { 266 // If one or the other blocks were not in a loop, the new block is not 267 // either, and thus LI doesn't need to be updated. 268 if (Loop *DestLoop = LI->getLoopFor(DestBB)) { 269 if (TIL == DestLoop) { 270 // Both in the same loop, the NewBB joins loop. 271 DestLoop->addBasicBlockToLoop(NewBB, *LI); 272 } else if (TIL->contains(DestLoop)) { 273 // Edge from an outer loop to an inner loop. Add to the outer loop. 274 TIL->addBasicBlockToLoop(NewBB, *LI); 275 } else if (DestLoop->contains(TIL)) { 276 // Edge from an inner loop to an outer loop. Add to the outer loop. 277 DestLoop->addBasicBlockToLoop(NewBB, *LI); 278 } else { 279 // Edge from two loops with no containment relation. Because these 280 // are natural loops, we know that the destination block must be the 281 // header of its loop (adding a branch into a loop elsewhere would 282 // create an irreducible loop). 283 assert(DestLoop->getHeader() == DestBB && 284 "Should not create irreducible loops!"); 285 if (Loop *P = DestLoop->getParentLoop()) 286 P->addBasicBlockToLoop(NewBB, *LI); 287 } 288 } 289 290 // If TIBB is in a loop and DestBB is outside of that loop, we may need 291 // to update LoopSimplify form and LCSSA form. 292 if (!TIL->contains(DestBB)) { 293 assert(!TIL->contains(NewBB) && 294 "Split point for loop exit is contained in loop!"); 295 296 // Update LCSSA form in the newly created exit block. 297 if (Options.PreserveLCSSA) { 298 // If > 1 identical edges to be split, we need to introduce the same 299 // number of the incoming blocks for the new PHINode. 300 createPHIsForSplitLoopExit( 301 SmallVector<BasicBlock *, 4>(NumSplitIdenticalEdges, TIBB), NewBB, 302 DestBB); 303 } 304 305 if (!LoopPreds.empty()) { 306 assert(!DestBB->isEHPad() && "We don't split edges to EH pads!"); 307 BasicBlock *NewExitBB = SplitBlockPredecessors( 308 DestBB, LoopPreds, "split", DT, LI, MSSAU, Options.PreserveLCSSA); 309 if (Options.PreserveLCSSA) 310 createPHIsForSplitLoopExit(LoopPreds, NewExitBB, DestBB); 311 } 312 } 313 } 314 } 315 316 return NewBB; 317 } 318 319 // Return the unique indirectbr predecessor of a block. This may return null 320 // even if such a predecessor exists, if it's not useful for splitting. 321 // If a predecessor is found, OtherPreds will contain all other (non-indirectbr) 322 // predecessors of BB. 323 static BasicBlock * 324 findIBRPredecessor(BasicBlock *BB, SmallVectorImpl<BasicBlock *> &OtherPreds) { 325 // Verify we have exactly one IBR predecessor. 326 // Conservatively bail out if one of the other predecessors is not a "regular" 327 // terminator (that is, not a switch or a br). 328 BasicBlock *IBB = nullptr; 329 for (BasicBlock *PredBB : predecessors(BB)) { 330 Instruction *PredTerm = PredBB->getTerminator(); 331 switch (PredTerm->getOpcode()) { 332 case Instruction::IndirectBr: 333 if (IBB) 334 return nullptr; 335 IBB = PredBB; 336 break; 337 case Instruction::Br: 338 case Instruction::Switch: 339 OtherPreds.push_back(PredBB); 340 continue; 341 default: 342 return nullptr; 343 } 344 } 345 346 return IBB; 347 } 348 349 bool llvm::SplitIndirectBrCriticalEdges(Function &F, 350 bool IgnoreBlocksWithoutPHI, 351 BranchProbabilityInfo *BPI, 352 BlockFrequencyInfo *BFI) { 353 // Check whether the function has any indirectbrs, and collect which blocks 354 // they may jump to. Since most functions don't have indirect branches, 355 // this lowers the common case's overhead to O(Blocks) instead of O(Edges). 356 SmallSetVector<BasicBlock *, 16> Targets; 357 for (auto &BB : F) { 358 if (isa<IndirectBrInst>(BB.getTerminator())) 359 Targets.insert_range(successors(&BB)); 360 } 361 362 if (Targets.empty()) 363 return false; 364 365 bool ShouldUpdateAnalysis = BPI && BFI; 366 bool Changed = false; 367 for (BasicBlock *Target : Targets) { 368 if (IgnoreBlocksWithoutPHI && Target->phis().empty()) 369 continue; 370 371 SmallVector<BasicBlock *, 16> OtherPreds; 372 BasicBlock *IBRPred = findIBRPredecessor(Target, OtherPreds); 373 // If we did not found an indirectbr, or the indirectbr is the only 374 // incoming edge, this isn't the kind of edge we're looking for. 375 if (!IBRPred || OtherPreds.empty()) 376 continue; 377 378 // Don't even think about ehpads/landingpads. 379 auto FirstNonPHIIt = Target->getFirstNonPHIIt(); 380 if (FirstNonPHIIt->isEHPad() || Target->isLandingPad()) 381 continue; 382 383 // Remember edge probabilities if needed. 384 SmallVector<BranchProbability, 4> EdgeProbabilities; 385 if (ShouldUpdateAnalysis) { 386 EdgeProbabilities.reserve(Target->getTerminator()->getNumSuccessors()); 387 for (unsigned I = 0, E = Target->getTerminator()->getNumSuccessors(); 388 I < E; ++I) 389 EdgeProbabilities.emplace_back(BPI->getEdgeProbability(Target, I)); 390 BPI->eraseBlock(Target); 391 } 392 393 BasicBlock *BodyBlock = Target->splitBasicBlock(FirstNonPHIIt, ".split"); 394 if (ShouldUpdateAnalysis) { 395 // Copy the BFI/BPI from Target to BodyBlock. 396 BPI->setEdgeProbability(BodyBlock, EdgeProbabilities); 397 BFI->setBlockFreq(BodyBlock, BFI->getBlockFreq(Target)); 398 } 399 // It's possible Target was its own successor through an indirectbr. 400 // In this case, the indirectbr now comes from BodyBlock. 401 if (IBRPred == Target) 402 IBRPred = BodyBlock; 403 404 // At this point Target only has PHIs, and BodyBlock has the rest of the 405 // block's body. Create a copy of Target that will be used by the "direct" 406 // preds. 407 ValueToValueMapTy VMap; 408 BasicBlock *DirectSucc = CloneBasicBlock(Target, VMap, ".clone", &F); 409 if (!VMap.AtomMap.empty()) 410 for (Instruction &I : *DirectSucc) 411 RemapSourceAtom(&I, VMap); 412 413 BlockFrequency BlockFreqForDirectSucc; 414 for (BasicBlock *Pred : OtherPreds) { 415 // If the target is a loop to itself, then the terminator of the split 416 // block (BodyBlock) needs to be updated. 417 BasicBlock *Src = Pred != Target ? Pred : BodyBlock; 418 Src->getTerminator()->replaceUsesOfWith(Target, DirectSucc); 419 if (ShouldUpdateAnalysis) 420 BlockFreqForDirectSucc += BFI->getBlockFreq(Src) * 421 BPI->getEdgeProbability(Src, DirectSucc); 422 } 423 if (ShouldUpdateAnalysis) { 424 BFI->setBlockFreq(DirectSucc, BlockFreqForDirectSucc); 425 BlockFrequency NewBlockFreqForTarget = 426 BFI->getBlockFreq(Target) - BlockFreqForDirectSucc; 427 BFI->setBlockFreq(Target, NewBlockFreqForTarget); 428 } 429 430 // Ok, now fix up the PHIs. We know the two blocks only have PHIs, and that 431 // they are clones, so the number of PHIs are the same. 432 // (a) Remove the edge coming from IBRPred from the "Direct" PHI 433 // (b) Leave that as the only edge in the "Indirect" PHI. 434 // (c) Merge the two in the body block. 435 BasicBlock::iterator Indirect = Target->begin(), 436 End = Target->getFirstNonPHIIt(); 437 BasicBlock::iterator Direct = DirectSucc->begin(); 438 BasicBlock::iterator MergeInsert = BodyBlock->getFirstInsertionPt(); 439 440 assert(&*End == Target->getTerminator() && 441 "Block was expected to only contain PHIs"); 442 443 while (Indirect != End) { 444 PHINode *DirPHI = cast<PHINode>(Direct); 445 PHINode *IndPHI = cast<PHINode>(Indirect); 446 BasicBlock::iterator InsertPt = Indirect; 447 448 // Now, clean up - the direct block shouldn't get the indirect value, 449 // and vice versa. 450 DirPHI->removeIncomingValue(IBRPred); 451 Direct++; 452 453 // Advance the pointer here, to avoid invalidation issues when the old 454 // PHI is erased. 455 Indirect++; 456 457 PHINode *NewIndPHI = PHINode::Create(IndPHI->getType(), 1, "ind", InsertPt); 458 NewIndPHI->addIncoming(IndPHI->getIncomingValueForBlock(IBRPred), 459 IBRPred); 460 NewIndPHI->setDebugLoc(IndPHI->getDebugLoc()); 461 462 // Create a PHI in the body block, to merge the direct and indirect 463 // predecessors. 464 PHINode *MergePHI = PHINode::Create(IndPHI->getType(), 2, "merge"); 465 MergePHI->insertBefore(MergeInsert); 466 MergePHI->addIncoming(NewIndPHI, Target); 467 MergePHI->addIncoming(DirPHI, DirectSucc); 468 MergePHI->applyMergedLocation(DirPHI->getDebugLoc(), 469 IndPHI->getDebugLoc()); 470 471 IndPHI->replaceAllUsesWith(MergePHI); 472 IndPHI->eraseFromParent(); 473 } 474 475 Changed = true; 476 } 477 478 return Changed; 479 } 480